Google talks DNN at GTC2015

March 2015 – At GTC 2015, the opening keynote by Jen-Hsun Huang, talked about Deep Neural Networks as a task that is a best fit for GPU compute capabilities. Following this introduction as a next generation application, Jeff Dean gave a keynote on Large Scale Deep Learning. These learning systems are well suited to GPU compute analysis due to their parallel architecture for the algorithms and for a parallel architecture for data.

The learning systems are implementations of neural nets that are different depths. One of the common points for the machine learning is it is best done with lots of data. The algorithms to learn not only need the items and interest, but the context of those items. Luckily on the web there is huge amounts of data in text, audio, images, activities and situations to gather the learning from. The new GPUs allow for following the basic premise of the DNN (deep neural nets) which humans can make decisions typically in 0.1sec, which is the same amount of time it takes for 10-levels of Neural Net Logic.

Google\’s Jeff Dean gives Keynote at GTC 2015

These techniques are being used to identify objects in images, speech in sound files, commands and objects in text, emotion from text and audio. The DNN methods produce more reliable results in a fast time than standard computational methods. This will allow for better voice recognition, search on video that has no metadata, search on images that have no description or only geotags. As more information on the web moves from text based to media based, the ability to search and extract information from these sources, as well as social media, the amount of currently searchable content is dropping. The anticipation that over 80% of web traffic and records will be media in a few years, means these identification technologies are critical to being able to data-mine content on the web.